LLM API Pricing Compared: What Inference Actually Costs per Million Tokens

LLM API pricing compared across OpenAI, Anthropic, and Google: the same job costs 18 cents on the cheapest model and $20 on a flagship — verified rates.

By Galactic News Editorial · Published 2026-11-09T07:37:54+00:00 · Updated 2026-11-09T07:37:54+00:00

LLM API Pricing Compared: What Inference Actually Costs per Million Tokens

LLM APIs bill per million tokens, and the price spread between the cheapest and the most expensive model is more than 100 times. This guide compares the current per-token rates from OpenAI, Anthropic, and Google, using each vendor's own pricing page as the source. All prices are in US dollars per million tokens, and they change often, so confirm them on the official page before you set a budget. The rates below were read from the three official pricing pages on 9 November 2026.

LLM API Pricing Compared: The Three Vendors at a Glance

Two price rises are already scheduled, so read the dates carefully before you budget. Gemini 4 Argon costs $2 for input and $10 for output through 30 November 2026, then doubles to $4 and $20 from 1 December 2026. Gemini 3.8 Flash costs 75 cents and $3.75 through 31 December 2026, then doubles to $1.50 and $7.50 from 1 January 2027. Gemini 2.5 Flash is deprecated, and from 1 December 2026 its API requests will route automatically to Gemini 3.5 Flash-Lite and be billed at that model's rates. Budget at the future price if your project will still be running after those dates.

OpenAI: Four Tiers From 10 Cents to $50

OpenAI prices its models on a wide ladder, with a clear step at each rung. GPT-6 Luna is the cheapest at 10 cents per million input tokens and 50 cents per million output tokens. GPT-6.1 Terra sits at $1 input and $5 output, while GPT-6.1 Sol, the workhorse of the lineup, costs $2 input and $10 output. The flagship GPT-6 Astra costs $10 input and $50 output, and a specialist cyber model, GPT-6 Cyber, runs higher still at $20 input and $100 output. Cached input is far cheaper than fresh input on every model, and the discount is steepest at the top of the range. Astra's cached input costs $1 per million, which is one tenth of its fresh input price, and Sol's cached input costs 10 cents. That makes prompt caching the single biggest lever on an OpenAI bill, because repeated context can cut input costs by 90 percent. OpenAI also lists promotional pricing for GPT-5.6 Sol until 30 November 2026, so check the pricing page for the current promotional dates before you rely on them.

Anthropic: Four Models, One Price Ladder

Anthropic's Claude lineup has four rungs, and each one has a clear job. Claude Haiku 5.5 costs 10 cents per million input tokens and 50 cents per million output tokens, but only for prompts up to 100,000 tokens. Above 100,000 tokens, Haiku costs 50 cents input and $2.50 output, so long prompts jump fivefold in price. Claude Sonnet 5.5 costs $2 input and $10 output, and Anthropic presents it as the best combination of speed and intelligence for everyday coding and writing. Claude Opus 5.5 costs $4 input and $20 output, and it is the top standard model for long-running agentic work. Claude Fable 5.1 sits above them all at $10 input and $50 output, for demanding reasoning and long-horizon agentic tasks. Prompt caching is cheap across the lineup, since a cache read costs one tenth of the standard input price on most models, and only 5 percent on Opus 5.5 and Sonnet 5.5. The Batch API halves input and output prices for work that can wait for asynchronous processing, which is a straightforward 50 percent saving. For most workloads, the choice is simple: use Haiku for high-volume simple work, Sonnet for everyday agents, Opus when the task genuinely needs it, and Fable only for the hardest reasoning jobs.

Google Gemini: The Cheapest Floor, With Price Rises Coming

Google's Gemini API has the widest range of any vendor here, from 10-cent input tokens to a flagship that doubles in price next month. Gemini 2.5 Flash-Lite is the cheapest model at 10 cents input and 40 cents output per million tokens, which makes it the floor for high-volume work. Gemini 2.5 Flash costs 30 cents input and $2.50 output, but it is deprecated, and its traffic moves to Gemini 3.5 Flash-Lite from 1 December 2026. Gemini 3.8 Flash costs 75 cents input and $3.75 output at its introductory price, and that price doubles on 1 January 2027 to $1.50 input and $7.50 output. Gemini 4 Argon, the flagship reasoning model, costs $2 input and $10 output through 30 November 2026, then doubles to $4 and $20 from 1 December 2026. Google offers a free tier on several models for experimentation, with free input and output tokens inside its rate limits, which makes Gemini the easiest vendor to prototype with. The Batch API cuts standard prices by 50 percent, the same discount Anthropic offers. The lesson from Google's page is to watch the calendar: two of its four current models double in price within weeks of this guide being written, so a budget built on today's rates will be wrong by January.

What One Million Tokens Actually Costs: A Worked Example

Price tables hide the real question, which is what a job costs. Take a job with 1 million input tokens and 200,000 output tokens, which is roughly 750,000 words of input text, so this is a large document job rather than a single chat message. Output tokens cost about five times input tokens on most models, so output dominates the bill even though there are fewer of them. Here is what that job costs on each model, at the standard rates on the pricing pages.

• Sol: $2 input plus $2 output (200,000 tokens at $10 per million). Total: $4.
• Astra: $10 input plus $10 output. Total: $20.
• Luna: 10 cents input plus 10 cents output. Total: 20 cents.
• Sonnet: $2 input plus $2 output. Total: $4.
• Opus: $4 input plus $4 output. Total: $8.
• Haiku: 10 cents input plus 10 cents output. Total: 20 cents.
• Fable: $10 input plus $10 output. Total: $20.
• Argon: $2 input plus $2 output. Total: $4.
• 3.8 Flash: 75 cents input plus 75 cents output. Total: $1.50.
• Flash-Lite: 10 cents input plus 8 cents output. Total: 18 cents.

The same job costs 18 cents on the cheapest model and $20 on a flagship, which is more than a 100-fold difference. The right model is the cheapest one that handles the job well, so start cheap and upgrade only when quality fails. Note that the Argon and 3.8 Flash totals above use the prices in force on 9 November 2026, and both will cost twice as much after their scheduled rises.

Free Tiers and Trial Credits: Test Before You Pay

Every vendor offers a way to test before you commit, but the offers are very different. Google provides a free tier on several Gemini models, including Flash, Flash-Lite, and 2.5 Pro, with input and output free of charge within the free tier's rate limits. That makes Gemini the easiest vendor to prototype with, since you can run real traffic without a card. Anthropic gives new API users a small amount of free test credit, and you must add a payment method before you call the API, since usage is billed per token. OpenAI's API has no published free plan, so you pay from the first token. ChatGPT subscriptions are separate from API billing, and paying for ChatGPT Plus does not include API credit. The practical advice is simple: prototype on Google's free tier or Anthropic's test credit, measure your real token usage for a week, and then pick a vendor based on measured cost rather than the price table. Token counts surprise everyone in the first month, because a chat feature that feels small can burn millions of tokens a day once users arrive.

Four Ways to Cut Your Inference Bill

1. Step 1: Cache repeated context. Cached input costs about one tenth of fresh input on OpenAI and Anthropic, and as little as 5 percent on Opus 5.5 and Sonnet 5.5. Put system prompts and long documents in the cache, and the savings start on the second call.
2. Step 2: Use batch processing. The Anthropic Batch API halves prices, and Google's batch discount is 50 percent as well. Use batch for work that does not need an instant answer, such as nightly summarisation or bulk classification.
3. Step 3: Route by task. Send classification and extraction to a cheap model such as Luna or Flash-Lite, and reserve the flagship for reasoning and long-horizon work. Most requests in a typical app do not need the most expensive model.
4. Step 4: Watch the long-prompt step. Haiku jumps fivefold above 100,000 tokens, and Gemini's flagship prices double on their scheduled dates. Split long prompts, or switch models before the threshold bites.

Set a monthly budget alert in your cloud billing console, because token usage grows quietly. A prompt that gets longer by 10 percent a month doubles your bill in about seven months, and by then the habit is hard to break.

Bottom Line

LLM API pricing compared across the three vendors shows the same pattern. Cheap models cost about 10 cents per million input tokens, workhorse models cost about $2, and flagships cost $10 or more. Output tokens cost roughly five times input tokens, so short answers save real money. Prompt caching and batch processing cut costs by half or more, and they are the two cheapest wins available. Two Gemini prices double within weeks, so plan for that now rather than discovering it in a January invoice. Start on the cheapest model that works, measure your real usage for a week, and upgrade only when quality proves the need.